Research On Face Recognition Under Unconstrained Conditions Based On Deep Learning

With the development of society,face recognition as one of biometric identification technology has been widely used in video monitoring,phone unlocking,unmanned retail,et al.Compare with iris recognition and fingerprint identification,face recognition has more advantages in non-contact and remote identification.However,because of the illumination,occlusion,pose,expression and et al,unconstrained face recognition is a very important yet extremely challenging problem.In recent years,deep learning techniques have significantly advanced large-scale unconstrained face recognition,arguably driven by rapidly increasing resource of face images.But,it is cost prohibitive and time consuming to structure a large human face dataset s with correctly annotated human facial images in unconstrained environment.To this end,it is a burning problem to further improve the accuracy of unconstrained face recognition by using the constrained face dataset.In order to solve this problem,the two types of modifications for training samples and loss functions are proposed respectively in this paper as following:(1)For training samples.There are two methods of data augmentation are proposed: a.3D facial deformation model,which can produce t he multiple face images with different poses by inputting one facial image.b.Facial landmarks detection based eyeglass facial images synthesize.After employing this two approaches,the unconstrained face data can be augmented drastically.(2)For loss function.We proposed a modified Softmax loss function,namely Scalable Softmax loss function,which can increase the punishment of hard samples while the decrease the punishment of easy samples,simultaneously.As the result,the robustness of networks ca n be further enhanced.Experimental results on CASIA-WebFace,MS-Celeb-1M,LFW,YTF and our self-build dataset GS-Face demonstrate that our method can improve the performance of unconstrained face recognition algorithm effectively.